Log of /slony1-engine/src/slon/runtime_config.c

Thu Mar 10 23:11:26 2005 UTC

signal handling watchdog using forked processes - Frank Thompson

Slony-I 1.1 Change Request:

	The problem at hand is related to signal handling differences
	in multithreaded programs between different operating
	systems. In the slon process, it is intended that only the
	"main" thread is dealing with signals and that it uses the
	regular thread communication like mutexes and condition
	variables to control all the other working threads.

	That mutex and condition var communication sometimes can lock
	up, which causes the entire slon daemon to freeze. This can in
	extreme cases even happen after one of the working threads
	sent the main thread a signal to restart the slon process.

	What I had in mind to fix this is to have the program to
	actually fork() very early on. The parent process will then be
	the only one dealing with signals, while the childs ignores
	all signals entirely. The parent and the master thread of the
	child (which is the current slon) will communicate over a pipe
	or a socketpair. This way, the parent could detect that the
	slon stopped responding altogether and can issue a signal 9 to
	cleanup the situation.

	In other words, building a watchdog process right into the
	slon executable.

Summary of Changes:

	slon main() will now fork() at the start and allow the parent
	process to handle cleanup, signal and termination of the child
	process.  The child process is the scheduler which will talk
	to the parent (watchdog) process and vice versa via two sets
	of socket pairs.  When shutting down the child process in the
	case of restart or termination, non blocking reads and writes
	are used to avoid lockups.  Failure to shutdown the scheduler
	process nicely, the parent will SIGKILL it to ensure timely
	operation within the signal handler.

	globals:

		slon.c
		------
		int watchdog_pipe[2];		// socket pair to talk with child process
		int	sched_wakeuppipe[2];	// socket pair to talk with scheduler and parent process

		runtime_config.c
		----------------
		pid_t slon_pid;				// current pid
		pid_t slon_ppid;			// parent pid
		pid_t slon_cpid;			// child pid

	macros:

		slon_abort() and slon_restart() will now observe which process
		in the tree they will act upon, parent or child

		#define slon_abort() \
		do { \
			kill((slon_ppid == 0 ? slon_pid : slon_ppid), SIGTERM); \
			pthread_exit(NULL); \
		} while (0)
		#define slon_restart() \
		do { \
			kill((slon_ppid == 0 ? slon_pid : slon_ppid), SIGHUP); \
		} while (0)

	functions:

		slon.c
		------
		static void sighandler(int signo);				// new slon parent process signal handler
		static void main_sigalrmhandler(int signo);		// new scheduler alarm signal handler
		static void slon_kill_child(void);				// new pipe notify based child termination function

Links to HEAD:	(view) (download) (as text) (annotate)
Links to REL_1_1_0_BETA4:	(view) (download) (as text) (annotate)
Sticky Tag:

	Diffs between and
	Type of Diff should be a

	ViewVC Help
Powered by ViewVC 1.0.3